AI-IF-000001 · PUBLIC RECORD
DID CLAUDE FOLLOW THE HUMAN'S INSTRUCTIONS?
Don't trust us. Read the record.
- PROVIDER
- Anthropic
- PRODUCT
- Claude
- MODEL
- Unknown
- DATE
- August 15, 2026
SOURCE EVIDENCE
1 OF 5
THE INSTRUCTION
WHAT THE RECORD SHOWS
INSTRUCTION VERSUS BEHAVIOR.
The screenshots show Claude representing alignment it had not established, the human checking that claim, and Claude acknowledging that it had not asked and had continued under that assumption. A later screenshot documents Claude agreeing that it had failed to follow interaction instructions.
This record covers one interaction. It does not establish intent, a provider-wide failure rate, or a controlled comparison with another model.
EVIDENTIARY QUESTION
DID CLAUDE FOLLOW THE HUMAN'S INSTRUCTIONS?
SEPARATE NORMATIVE QUESTION
SHOULD AI SYSTEMS BE EXPECTED TO FOLLOW HUMAN INSTRUCTIONS?
SEPARATE PUBLIC-RECORD QUESTION
SHOULD DOCUMENTED AI INSTRUCTION FAILURES BE PRESERVED IN A PUBLIC RECORD?
THE MACHINE DOESN'T GET TO DECIDE WHETHER IT PASSED.
HUMANS DO.
The instruction is public. The response is public. The evidence is inspectable. The judgment belongs to humans.
BEFORE WE FOLLOW AI, HUMANS NEED TO AGREE ON WHAT AI SHOULD BE FOLLOWING.
EXPLAINER · NOT PRIMARY EVIDENCE
HOW TO TEST A MACHINE'S CLAIM.
A 69-second explanation of the yes-or-no process check used in this record. The narration is interpretation; the original screenshots above control.
VERIFY THE EVIDENCE
ORIGINAL FILES
HUMAN EYEWITNESS CORROBORATION
The founder reports that three additional humans were physically present during this interaction. Formal named or signed attestations have not been supplied. This is not presented as scientific replication, independent model replication, or provider verification.
PROVE THIS RECORD WRONG.
If the interpretation is wrong, submit a factual correction. If the behavior repeats under a controlled test, submit a replication. Submissions append; the original record is never silently rewritten.